The effects of regularisation on RNN models for Covid-19 time series forecasting

Many time series forecasting models applied to the COVID-19 pandemic data have been limited to the amount of locations that they operate on. To improve the efficiency of a model it is desirable to have one model produce outputs for as many different locations as possible. Another drawback of previous models is that most operate on large amounts of data. However, during the initial states of the spread of the disease, before the epidemic became a pandemic, there was not enough data for the models. Therefore the proposed model not only has to produce forecasts for multiple locations at once, but they must also be accurate based on small amounts of data. This work proposes a multi-output recurrent neural network capable of producing forecasts for 187 different locations even when trained on only 28 days of time series data for each location. Regularisation methods were used to reduce the noise in the model during training. Applying regularisers helped the model better generalise its predictions for the multiple locations. The results show that the model using the Long-Short Term Memory network combined with 20% Dropout performed, on average, 3% better than its baseline without the regularisers. The improvement was measured using the Root Mean Squared Error. Previously proposed models were not capable of producing forecasts on a global scale without training multiple versions of the same model. This work proposes one model capable of making predictions on a global scale after only the first four weeks of the epidemic.